Six studies in the stock check the bibliographic references that models generate: in four the models were asked for reference lists (Naser, GhostCite, the eight chatbots, Chelli), in two they drafted text with references (medical research introductions in Omar; letters to the editor and original articles in Ozbek). Three state how the references were verified (two by automated matching against bibliographic databases, one by hand), three are observed as abstracts that do not state who verified: invalid or fabricated shares run from 11.4 to 56.8 percent (ten commercial LLMs), 14.23 to 94.93 percent (thirteen LLMs), 28.6 to 91.4 percent (GPT-3.5, GPT-4, Bard), and 39.8 percent of 400 references from eight free chatbots. All of them measure whether the reference exists and is bibliographically correct; none checks whether an existing reference supports the claim it is attached to, so none of these rates is a value of the quantity the question asks for.

Naser and GhostCite contribute the two large automated audits and the ranges; the eight-chatbot study contributes a hand-checked sample; Chelli contributes a comparison against the reference lists of published systematic reviews under a rule of two wrong fields out of three, from an abstract that does not state who checked; Omar contributes a second medical comparison in which no model is free of fabricated references (correctness 77.2 and 54.0 percent); Ozbek contributes that a composite score mixing existence with topical relevance is still not a support check. Existence is a precondition of support: a fabricated reference supports nothing, so in the four reference-list studies the complements of these rates are loose upper bounds on support (a reference that does not exist supports nothing; one that exists may or may not), but the rates themselves bound nothing. Scope: only the Naser models are stated to have run without retrieval; the GhostCite rates pool runs with and without online search, which showed no consistent effect, and the other parents do not state the retrieval status. No parent states that its models cited a retrieved page, and two of the parents with unstated retrieval status include Perplexity (Ozbek, the eight chatbots), so the step does not claim that retrieval was absent.

One of these studies turns out to have checked the content of the cited work against the claim, which would make its rate a support rate; or a study shows that for retrieval-backed assistants existence failures and support failures are the same citations. Query: methods sections of the six papers for any content check.

AI makes up 40 percent of its citations. Too coarse: that figure comes from studies that asked chatbots for reference lists and checked whether the references exist; it says nothing about whether the citations of a searching assistant support its claims.

Attacker run 3 (attacker-grok-4.7, x-attack-step): applied, step wording corrected in place; the rates do not bound support from above, their complements do. Conclusion unchanged.